On Thursday, 15 January 2026 19:43:03 CET Janusz Krzysztofik wrote:
> The smem-oom subtest can expectedly result in oom-killer being triggered,
> which then dumps a call trace from a process that triggered it. If that
> happens to be a process that executes drm or i915 functions then the call
> trace dump contains lines recognized by igt_runner running in piglit mode
> as potential warnings. If severity of the call trace dump messages is
> NOTICE or higher, which isn't unlikely, then a dmesg-warn result is
> reported despite successful completion of the subtest.
>
> Fortunately, severity of those call trace dump messages depends on kernel
> default log level which can be controlled from user space over sysctl.
>
> To avoid false failure reports, relax kernel default log level to INFO so
> those log lines are ignored by igt_runner in piglit mode at an expense of
> call traces from real issues potentially detected by the subtest not
> contributing to the igt_runner reported result. Since those call traces
> are still available to developers, only submitted with reduced severity,
> that shouldn't hurt as long as the igt_runner still abandons further
> execution and reports an abort result on a kernel taint.
>
> v4: In the exit handler, restore the default log level with a simple
> write() from a formerly prepared buffer, and abort verbosely if that
> doesn't succeed (Kamil).
> v3: Move cleanup to an exit handler in case we are killed (Kamil).
> v2: Move default log level setup inside subtest smem-oom (Kamil),
> - move cleanup there as well.
>
> Closes: https://gitlab.freedesktop.org/drm/i915/kernel/-/issues/5493
> Cc: Kamil Konieczny <[email protected]>
> Signed-off-by: Janusz Krzysztofik <[email protected]>
> ---
> tests/intel/gem_lmem_swapping.c | 91 ++++++++++++++++++++++++++++++++-
> 1 file changed, 89 insertions(+), 2 deletions(-)
>
> diff --git a/tests/intel/gem_lmem_swapping.c b/tests/intel/
gem_lmem_swapping.c
> index adae26716c..1b07124aff 100644
> --- a/tests/intel/gem_lmem_swapping.c
> +++ b/tests/intel/gem_lmem_swapping.c
> @@ -9,6 +9,7 @@
> #include "i915/intel_memory_region.h"
> #include "igt.h"
> #include "igt_kmod.h"
> +#include "runnercomms.h"
> #include <unistd.h>
> #include <stdlib.h>
> #include <stdint.h>
> @@ -661,6 +662,38 @@ static void gem_leak(int fd, uint64_t alloc)
> gem_madvise(fd, handle, I915_MADV_DONTNEED);
> }
>
> +static int printk = -1;
> +static char log_levels[4];
> +
> +static void printk_exit_handler(int sig)
> +{
> + char *message = NULL;
> + int n;
> +
> + if (printk < 0)
> + return;
> +
> + n = strlen(log_levels);
> + if (!n) {
> + message = strdup("no defaults saved, unable to restore /
proc/sys/kernel/printk\n");
> + } else if (write(printk, log_levels, n) != n) {
> + n = errno;
> + if (asprintf(&message,
> + "restoring /proc/sys/kernel/printk defaults
failed, errno: %d\n",
> + n) < 0)
> + message = NULL;
Intended to prevent from processing an undefined pointer, that's incorrect,
because that effectively disables the abort path despite the unsuccessful
restore of log levels. Now I think I should rather reserve a local 80-char
buffer on stack and sprintf the message to that buffer. Let me submit v5.
Thanks,
Janusz
> + }
> + close(printk);
> +
> + if (message) {
> + log_to_runner_sig_safe(message, strlen(message));
> + free(message);
> + abort();
> + }
> +
> + printk = -1;
> +}
> +
> static int *lmem_done;
>
> static void smem_oom_exit_handler(int sig)
> @@ -861,8 +894,62 @@ int igt_main_args("", long_options, help_str,
opt_handler, NULL)
> }
>
> igt_describe("Exercise local memory swapping during exhausting
system memory");
> - dynamic_lmem_subtest(region, regions, "smem-oom")
> - test_smem_oom(i915, ctx, region);
> + igt_subtest_with_dynamic("smem-oom") {
> + int console_log_level, default_log_level;
> + unsigned int fd, i = 0;
> + FILE *stream = NULL;
> +
> + /*
> + * This subtest can result in oom-killer being triggered,
which
> + * then dumps a call trace from a process that triggered
it.
> + * If that happens to be a process that executes drm or
i915
> + * functions then the call trace dump contains lines
recognized
> + * by igt_runner as warnings and a dmesg-warn result is
> + * reported. To avoid false failure reports, relax kernel
> + * default log level to INFO for those lines to be ignored
by
> + * igt_runner in piglit mode, at an expense of call traces
from
> + * potential real issues not contributing to the
igt_runner
> + * reported result. Since those call traces are still
available
> + * to developers, only displayed with relaxed severity,
that
> + * shouldn't hurt as long as igt_runner still abandons
further
> + * execution and reports an abort result on a kernel
taint.
> + */
> + fd = open("/proc/sys/kernel/printk", O_RDWR);
> + if (!igt_debug_on(fd < 0))
> + stream = fdopen(fd, "r+");
> +
> + if (igt_debug_on(!stream))
> + close(fd);
> + else
> + i = fscanf(stream, "%d %d", &console_log_level,
&default_log_level);
> +
> + if (igt_debug_on(i != 2) || default_log_level >= 6) {
> + i = 0;
> + fclose(stream);
> + } else {
> + i = snprintf(log_levels, sizeof(log_levels), "%d
%d",
> + console_log_level,
default_log_level);
> + }
> +
> + if (!igt_debug_on(i != 3))
> + printk = dup(fd);
> +
> + if (!igt_debug_on(printk < 0)) {
> + igt_install_exit_handler(printk_exit_handler);
> +
> + rewind(stream);
> + igt_debug_on(fprintf(stream, "%d 6",
console_log_level) != 6);
> + }
> +
> + for (i = 0; i < regions->num_regions; i++) {
> + region = ®ions->regions[i];
> + if (region->region.memory_class ==
I915_MEMORY_CLASS_DEVICE)
> + igt_dynamic_f("lmem%u", region-
>region.memory_instance)
> + test_smem_oom(i915, ctx,
region);
> + }
> +
> + printk_exit_handler(0);
> + }
>
> igt_fixture() {
> intel_allocator_multiprocess_stop();
>
